| 1 | /* |
| 2 | * GMC (Global Motion Compensation), AltiVec-enabled |
| 3 | * |
| 4 | * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> |
| 5 | * |
| 6 | * This file is part of FFmpeg. |
| 7 | * |
| 8 | * FFmpeg is free software; you can redistribute it and/or |
| 9 | * modify it under the terms of the GNU Lesser General Public |
| 10 | * License as published by the Free Software Foundation; either |
| 11 | * version 2.1 of the License, or (at your option) any later version. |
| 12 | * |
| 13 | * FFmpeg is distributed in the hope that it will be useful, |
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 16 | * Lesser General Public License for more details. |
| 17 | * |
| 18 | * You should have received a copy of the GNU Lesser General Public |
| 19 | * License along with FFmpeg; if not, write to the Free Software |
| 20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 21 | */ |
| 22 | |
| 23 | #include "libavutil/mem.h" |
| 24 | #include "libavutil/ppc/types_altivec.h" |
| 25 | #include "libavutil/ppc/util_altivec.h" |
| 26 | #include "libavcodec/mpegvideodsp.h" |
| 27 | |
| 28 | #if HAVE_ALTIVEC |
| 29 | /* AltiVec-enhanced gmc1. ATM this code assumes stride is a multiple of 8 |
| 30 | * to preserve proper dst alignment. */ |
| 31 | static void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, |
| 32 | int stride, int h, int x16, int y16, int rounder) |
| 33 | { |
| 34 | int i; |
| 35 | const DECLARE_ALIGNED(16, unsigned short, rounder_a) = rounder; |
| 36 | const DECLARE_ALIGNED(16, unsigned short, ABCD)[8] = { |
| 37 | (16 - x16) * (16 - y16), /* A */ |
| 38 | (x16) * (16 - y16), /* B */ |
| 39 | (16 - x16) * (y16), /* C */ |
| 40 | (x16) * (y16), /* D */ |
| 41 | 0, 0, 0, 0 /* padding */ |
| 42 | }; |
| 43 | register const vector unsigned char vczero = |
| 44 | (const vector unsigned char) vec_splat_u8(0); |
| 45 | register const vector unsigned short vcsr8 = |
| 46 | (const vector unsigned short) vec_splat_u16(8); |
| 47 | register vector unsigned char dstv, dstv2, srcvB, srcvC, srcvD; |
| 48 | register vector unsigned short tempB, tempC, tempD; |
| 49 | unsigned long dst_odd = (unsigned long) dst & 0x0000000F; |
| 50 | unsigned long src_really_odd = (unsigned long) src & 0x0000000F; |
| 51 | register vector unsigned short tempA = |
| 52 | vec_ld(0, (const unsigned short *) ABCD); |
| 53 | register vector unsigned short Av = vec_splat(tempA, 0); |
| 54 | register vector unsigned short Bv = vec_splat(tempA, 1); |
| 55 | register vector unsigned short Cv = vec_splat(tempA, 2); |
| 56 | register vector unsigned short Dv = vec_splat(tempA, 3); |
| 57 | register vector unsigned short rounderV = |
| 58 | vec_splat((vec_u16) vec_lde(0, &rounder_a), 0); |
| 59 | |
| 60 | /* we'll be able to pick-up our 9 char elements at src from those |
| 61 | * 32 bytes we load the first batch here, as inside the loop we can |
| 62 | * reuse 'src + stride' from one iteration as the 'src' of the next. */ |
| 63 | register vector unsigned char src_0 = vec_ld(0, src); |
| 64 | register vector unsigned char src_1 = vec_ld(16, src); |
| 65 | register vector unsigned char srcvA = vec_perm(src_0, src_1, |
| 66 | vec_lvsl(0, src)); |
| 67 | |
| 68 | if (src_really_odd != 0x0000000F) |
| 69 | /* If (src & 0xF) == 0xF, then (src + 1) is properly aligned |
| 70 | * on the second vector. */ |
| 71 | srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src)); |
| 72 | else |
| 73 | srcvB = src_1; |
| 74 | srcvA = vec_mergeh(vczero, srcvA); |
| 75 | srcvB = vec_mergeh(vczero, srcvB); |
| 76 | |
| 77 | for (i = 0; i < h; i++) { |
| 78 | dst_odd = (unsigned long) dst & 0x0000000F; |
| 79 | src_really_odd = (((unsigned long) src) + stride) & 0x0000000F; |
| 80 | |
| 81 | dstv = vec_ld(0, dst); |
| 82 | |
| 83 | /* We'll be able to pick-up our 9 char elements at src + stride from |
| 84 | * those 32 bytes then reuse the resulting 2 vectors srvcC and srcvD |
| 85 | * as the next srcvA and srcvB. */ |
| 86 | src_0 = vec_ld(stride + 0, src); |
| 87 | src_1 = vec_ld(stride + 16, src); |
| 88 | srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src)); |
| 89 | |
| 90 | if (src_really_odd != 0x0000000F) |
| 91 | /* If (src & 0xF) == 0xF, then (src + 1) is properly aligned |
| 92 | * on the second vector. */ |
| 93 | srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src)); |
| 94 | else |
| 95 | srcvD = src_1; |
| 96 | |
| 97 | srcvC = vec_mergeh(vczero, srcvC); |
| 98 | srcvD = vec_mergeh(vczero, srcvD); |
| 99 | |
| 100 | /* OK, now we (finally) do the math :-) |
| 101 | * Those four instructions replace 32 int muls & 32 int adds. |
| 102 | * Isn't AltiVec nice? */ |
| 103 | tempA = vec_mladd((vector unsigned short) srcvA, Av, rounderV); |
| 104 | tempB = vec_mladd((vector unsigned short) srcvB, Bv, tempA); |
| 105 | tempC = vec_mladd((vector unsigned short) srcvC, Cv, tempB); |
| 106 | tempD = vec_mladd((vector unsigned short) srcvD, Dv, tempC); |
| 107 | |
| 108 | srcvA = srcvC; |
| 109 | srcvB = srcvD; |
| 110 | |
| 111 | tempD = vec_sr(tempD, vcsr8); |
| 112 | |
| 113 | dstv2 = vec_pack(tempD, (vector unsigned short) vczero); |
| 114 | |
| 115 | if (dst_odd) |
| 116 | dstv2 = vec_perm(dstv, dstv2, vcprm(0, 1, s0, s1)); |
| 117 | else |
| 118 | dstv2 = vec_perm(dstv, dstv2, vcprm(s0, s1, 2, 3)); |
| 119 | |
| 120 | vec_st(dstv2, 0, dst); |
| 121 | |
| 122 | dst += stride; |
| 123 | src += stride; |
| 124 | } |
| 125 | } |
| 126 | #endif /* HAVE_ALTIVEC */ |
| 127 | |
| 128 | av_cold void ff_mpegvideodsp_init_ppc(MpegVideoDSPContext *c) |
| 129 | { |
| 130 | #if HAVE_ALTIVEC |
| 131 | c->gmc1 = gmc1_altivec; |
| 132 | #endif /* HAVE_ALTIVEC */ |
| 133 | } |